MarkItDown vs SkillClaw

MarkItDown is Microsoft's open-source Python utility that converts virtually any file format into clean, LLM-friendly Markdown. It handles PDFs, Word documents, PowerPoint presentations, Excel spreadsheets, HTML, CSV, JSON, XML, ZIP archives, images (with optional vision model descriptions), audio files (with transcription), YouTube URLs, and EPub files in one consistent interface. The key design philosophy is LLM-first: rather than trying to reproduce original formatting for human readers, MarkItDown preserves document structure—headings, lists, tables, links—in a format that language models naturally parse efficiently. It integrates with OpenAI-compatible vision clients for image descriptions and supports speech transcription for audio content. With 108k+ GitHub stars and still gaining nearly 2,000 per day, MarkItDown has become the default document ingestion layer for countless AI pipelines. As agents increasingly need to process real-world enterprise documents, this kind of robust conversion utility becomes critical infrastructure—turning messy business files into clean inputs that Claude or GPT-4o can reason about without token-wasting formatting artifacts.

SkillClaw is a research framework from Alibaba's AMAP-ML team that enables collective skill evolution for LLM agent systems deployed at scale. The core idea: instead of each user's agent interactions existing in isolation, SkillClaw aggregates anonymized skill-improvement signals across all users to continuously refine a shared library of reusable agent skills — without requiring centralized fine-tuning. The framework introduces a three-component architecture: a Skill Extractor that identifies and catalogs atomic capabilities from interactions, a Skill Evolver that proposes improvements based on aggregate feedback, and a Skill Selector that routes tasks to the best-available skill version per user context. Published on April 9 and hitting #1 on Hugging Face trending papers this week with 277 upvotes, the paper reports significant improvements over per-user baselines on complex multi-step agentic tasks. This matters especially for production agent deployments where cold-start problems are severe — a new user's agent immediately benefits from millions of prior interactions. It's a fundamentally different model of agent improvement than either fine-tuning (expensive, periodic) or RAG (retrieval-only, no learning).